Open source

The ASCII Factory

How a team’s everyday corrections become a benchmark, and a benchmark becomes an AI skill that gets it right the first time. Two minutes, narrated, from a real run.

Why steer, benchmark and test a skill

1. Steer

   People already correct AI in plain words. Each correction says what good looks like.

2. Check

   Each correction becomes a yes-or-no check on that case. Accepted cases are the benchmark, built from real work.

3. Optimize

   A teacher model writes new versions of the skill from the checks it fails.

4. Test

   Every version runs on every case, first try. It replaces the skill only if it passes more checks.

In this run, one person steered an /ascii skill toward solid, shaded drawings that stand on a shadow. Before the first optimization each drawing took two corrections; after it, none. The skill went from 16 of 31 checks to 46 of 50, and of 13 versions tried, 5 won and 8 lost: tests kept every worse version from shipping.